Our study area is the Saitama prefecture of Japan, which is located at the western side of the Greater Tokyo area. It covers an area of 3,798 km2 and has a population size of 7,237 thousand. Most parts of Saitama can be regarded as Tokyo suburbs, and Saitama’s urban area is constantly expanding due to immigration to the Greater Tokyo Area.
This study collects Global Land Survey (GLS) satellite image datasets for 2000, 2005 and 2010. GLS datasets have eight bands, within which band 8 (Panchromatic) has a resolution of 15 meters, and the other seven bands have resolutions of 30 meters each. All bands except band 8 are resampled into 15 meters and combined for LU classification.
This study classifies four LU categories (water, agriculture, forest and built-up) using the supervised classification algorithm provided in ERDAS IMAGINE V2016 (Hexgon Geospatial, U.S.).
Supervised LU classification algorithm needs a prior knowledge of the LU distribution in the study area. More specifically, a set of pixels in a specific area is selected as training samples and defined with LU categories manually. Then, the algorithm is trained and used to recognize the LU categories for pixels in other parts of the satellite image based on this manually defined samples.
In this study, the criteria for selecting the sub-areas, which is used for providing prior knowledge, is: 1) sub-area should cover all four target LU categories (forest, agriculture land, residential land and water, see Table 3.5 for the definition of four LU categories); 2) given agriculture lands includes many different kinds of crops, which usually show colors with subtle differences on the satellite images, sub-area should have various agriculture lands with considerable variation of colors. Based on this criteria, the selection was conducted by visually interpreting the Landsat satellite images and the corresponding Google Earth historical satellite images.
Within the selected sub-areas, a set of pixels with varying colors is chosen as training
samples. The set of pixels covers most of colors with subtle difference and accounts for approximately 1/30 of all the pixels in these areas. The reference LU categories of the training samples are defined by visual interpretation, i.e., by comparing the Landsat satellite images with Google Earth historical satellite images that have finer resolution with approximately 5 meter, and visually determine the LU categories based on the geographical object shown on Google Earth historical satellite images. In the training process, the computer system continuously makes comparison between a seed pixel in the training samples and the contiguous pixels based on the similarity of spectral information of each pixel. Once the LU category of a contiguous pixel is accepted, the contiguous pixel is included as a training sample. Then the computer system will move on to recognize the neighboring pixels of the newly accepted pixel until every pixel in the area is recognized.
This study uses Mahalanobis distance method as supervised decision rule for the clas-sification. It is one of the most popular method used for processing remote sensing data.
Mahalanobis distance method calculates the spectral distance between the measurement vector (contains the measures of each spectral band) for the contiguous pixel and the mean vector for each training samples. The equation for the Mahalanobis distance for a target contiguous pixel, which needs to be classified, is
D= (X−Mc)TCovc−1(X−Mc) (3.10) where D denotes Mahalanobis distance, c is a particular class (LU category), X is the measurement vector of target pixel, Mcis the mean vector of the training sample of class (LU category) c, Cov denotes the covariance matrix of the pixel in the training samples with class (LU category) of c. The target pixel is assigned to the LU category with the lowest D.
Table 3.4 presents the confusion matrices of LUC, and Figure 3.3 shows LU maps of the three years. This study excludes the LU transitions that have transition rates below 1% and are left with transition from forest to agriculture, transition from agriculture to
forest, and transition from agriculture to built-up as the modeling objects.
Figure 3.3: Actual LU maps in Saitama prefecture of Greater Tokyo Area for 2000, 2005 and 2010
Given the possibility of varying transitional rules being present among the three LU transition types, This study separately develops transition probability estimation models and CA models for each of them. Geographic features are derived based on spatial data collected from the Ministry of Infrastructure, Land and Tourism of Japan, except for LU enrichment, which is directly calculated from LU maps. Satellite image patches of a certain size are cropped from the satellite images and used as input for the conv-net or CDAE model. The input image size is determined based on previous evidence on the effect of neighborhood size and trial-and-error experiments. The image input size is 27×27 for conv-net and 81×81 for CDAE-net. In addition, the satellite input image
Table 3.4: Confusion matrices from 2000 to 2005 and from 2005 to 2010 2005
Built-up Forest Agri. Water body
2000
Built-up 4090795 (98.99%) 13672 (0.33%) 20508 (0.50%) 7522 (0.18%)
Forest 22787 (0.72%) 2693884 (85.10%) 435253 (13.75%) 13672 (0.43%)
Agri. 419344 (7.48%) 575959 (10.27%) 4601687 (82.06%) 10937 (0.20%)
Water body 478 (0.39%) 888 (0.72%) 683 (0.55%) 121397 (98.34%)
2010
Built-up Forest Agri. Water body
2005
Built-up 4601286 (98.30%) 33151 (0.71%) 37887 (0.81%) 8530 (0.18%)
Forest 20894 (0.66%) 2564788 (80.75%) 565762 (17.81%) 24863 (0.78%)
Agri. 605314 (11.99%) 345704 (6.85%) 4085922 (80.93%) 11839 (0.23%)
Water body 947 (0.77%) 591 (0.48%) 591 (0.48%) 121397 (98.28%)
Notes:
1. The confusion matrix is presented as num of cells (the percentage).
2. The percentage is calculated by (numt−numt−1)/numt−1 wherenumdenotes the number of cells andtdenotes the time.
Table 3.5: Definition of land use categories in Saitama prefecture for 2000, 2005 and 2010
LU category Description
Agriculture
Lands used for growing crops including wet paddy filed, dry paddy field, swamp paddy field, fields used for growing wheat, upland rice, veget-ables, fruits, tea, wax tree, paper mulberry, hemp palm, etc., as well as grassland and lawn
Forest Lands where perennial plants are densely distributed
Built-up Lands where residential buildings, commercial buildings, etc. are densely
distributed
Water body Areas including river and river bed, artificial lake, natural lake, pond,
fish farm, etc. where are filled with water for most of the time Notes:
The definitions of LU categories are based on the information provided by National Land Information Division, National and Regional Policy Bureau of Japan
(http://nlftp.mlit.go.jp/ksj/gml/codelist/LandUseCd-09.html)
has seven bands, excluding band 6 (thermal) due to its low spatial variation.
The LUC models are trained on the 2000 and 2005 datasets, validated on a subset of data for 2005 and 2010, and tested on the whole dataset for 2005 and 2010. To minimize the spatial autocorrelation between the validation set and the test set to facilitate an
unbiased model evaluation, the validation set is extracted from a sub-region of Saitama covering approximately 15% of the total area. In terms of sampling, previous studies have normally used random or stratified sampling to avoid the influence of spatial au-tocorrelation. However, the mini-batch learning criterion of a neural network naturally mitigates the influence of spatial autocorrelation to some extent. This study uses a boot-strap over-sampling strategy in this study. Specifically, a mini-batch of data is randomly sampled from the dataset with replacement, and samples belonging to negative and posi-tive labels have the same proportion within a mini-batch. Over-sampling could make the model prone to over-fitting. To address this issue, this study considers the fine-tuning of the Gaussian noise coefficient. More discussion on the over-sampling is provided by Batista et al. (2004).